Book a demo
Topics · REGULATORY

Asset Mgmt Model Risk Checklist | Moring

A working AI model risk checklist for asset management covers ten items: a complete model inventory, defined model scope, risk appetite statements, tiered materiality assessments, independent validation standards, monitoring thresholds, override registers, vendor model terms, retention of validation evidence, and committee escalation paths. Each item maps to supervisory expectations from SR 11-7 onward and EU AI Act transparency duties.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

A working AI model risk checklist for asset management covers ten items: a complete model inventory, defined model scope, risk appetite statements, tiered materiality assessments, independent validation standards, monitoring thresholds, override registers, vendor model terms, retention of validation evidence, and committee escalation paths. Each item maps to supervisory expectations from SR 11-7 onward and EU AI Act transparency duties.

How it works

A checklist earns its place only if each item tests something observable, so the following list ties every entry to evidence an examiner, an internal auditor, or a due diligence consultant could request. Dutch supervisory analysis of asset managers published in November 2025 distilled eight good practices for model risk management, and United States expectations from SR 11-7 through the more recent SR 26-2 supply the supervisory backbone; the checklist below merges both with EU AI Act transparency duties into items an asset management firm can verify quarterly.

Item one is the model inventory with discovery, not just declaration. The register should list every model influencing investment decisions, risk reporting, or client communications, including vendor-supplied analytics and machine learning features embedded in platforms, with owner, purpose, tier, and lifecycle status. Item two is a written model definition, the first of the Dutch good practices, so that teams cannot argue an optimization routine or a generative research assistant falls outside scope. Firms that skip the definition spend committee time relitigating boundary cases repeatedly.

Item three is risk appetite and key risk indicators expressed in measurable terms, such as maximum tolerable forecast error, override frequency thresholds, or drift limits per tier. Item four is a materiality assessment rubric that assigns each model a tier determining validation depth and monitoring intensity. Item five is independent validation: documented standards covering assumptions, performance testing, and explainability, performed by reviewers outside the model's development line, with revalidation triggers tied to material change rather than the calendar alone.

Item six is continuous monitoring with thresholds that alert in near real time, not annual attestations. Item seven is an override register capturing when investment staff depart from model output and why, which supervisory reviewers treat as a direct window into how models and humans actually interact. Item eight is vendor model governance: contract terms securing change notifications, transparency sufficient for registration and validation, and the same inventory entry requirements as internally built models, a distinct control area in supervisory thematic reviews of AI model risk.

Item nine is evidence retention. Validation reports, committee minutes, breach investigations, and revalidation records must persist in retrievable form for the period examination practice expects, because an unprovable control is indistinguishable from an absent one. Governance control research highlights self-generating evidence, where controls write their own audit trail as a byproduct of operation, as the practical standard modern platforms reach. Item ten is escalation and retirement: documented paths from breach to committee, and retirement standards that force decommissioning decisions instead of letting unused models linger unmonitored.

Two cross-cutting checks close the list. First, framework consolidation: the NIST AI Risk Management Framework functions of govern, map, measure, and manage should map visibly onto these ten items so that EU AI Act transparency duties, SR 11-7 lineage, and SR 26-2 expectations are answered from one evidence base. Second, operating rhythm: every item needs a review frequency, an accountable executive, and a committee reporting line. A checklist without cadence decays into shelf documentation, which is the outcome the entire exercise exists to prevent.

Frequently asked questions

What are the essential items on an AI model risk checklist for asset managers?

Ten items cover the discipline: a discovery-backed model inventory, a written model definition, measurable risk appetite and key risk indicators, a tiered materiality rubric, independent validation standards, continuous monitoring thresholds, an override register, vendor model contract terms, retained validation evidence, and documented escalation and retirement paths. The list merges the Dutch regulator's eight good practices for asset managers with United States expectations from SR 11-7 through SR 26-2 and EU AI Act transparency duties. Each item needs an accountable executive and a review frequency, because a checklist without cadence decays into shelf documentation.

How often should an asset management firm review its model risk checklist items?

Cadence varies by item volatility. Inventory completeness and vendor model registrations warrant quarterly review because vendor platforms can enable AI capabilities without notice. Monitoring thresholds and override registers operate continuously but merit quarterly attestation that alerts fire and breaches reach committees. Validation standards and risk appetite statements justify annual review with off-cycle updates after material change or new supervisory guidance, such as the transition from SR 11-7 toward SR 26-2 expectations. Every item should carry its frequency on the checklist itself so the second line can verify operation rather than assume it. In asset management, vendor platform changes and model risk guidance updates are the two events that force off-cycle checklist review.

What evidence proves each checklist item is actually operating?

Evidence differs by item but shares one trait: it is produced by the control, not assembled afterward. Inventory entries show discovery dates and automated scans; materiality rubrics show tier assignments with rationales; validation files show reviewer independence and assumption limitations; monitoring shows timestamped alerts and committee minutes reviewing each breach; override registers show dated, reasoned entries. Control research highlights self-generating evidence, where platforms write their own audit trail during operation, as the modern standard. United States expectations under SR 11-7 and EU AI Act transparency duties are answered from this same evidence base, not from policy prose. The principle applies across the asset management model risk inventory: an unprovable control is indistinguishable from an absent one.

How does the checklist handle vendor-supplied AI models in asset management?

Vendor models enter the checklist as full inventory members with no exemption. Each registration records the platform embedding the model, its purpose, and its tier, and contract terms must secure change notifications plus transparency sufficient for registration and validation. Supervisory thematic reviews of AI model risk identify third-party AI as a distinct control area precisely because firms otherwise assume vendor assurance replaces their own validation. The checklist item therefore tests both the contractual rights the firm holds and the internal review the firm performed, since accountability for model behavior in investment workflows stays with the asset manager regardless of the vendor.

Sources

  1. - federalreserve.gov
  2. - federalreserve.gov
  3. - artificialintelligenceact.eu